implicit

All recommendation models in implicit implement the implicit.RecommenderBase interface. This ensures a consistent API across different model implementations, allowing you to use the same methods for training models and generating recommendations regardless of whether you are using a CPU or GPU-based model.

Key capabilities provided by the interface include:

• Training the model on sparse matrices of user/item/confidence weights.
• Generating recommendations for specific users.
• Finding related items.

For x86_64 Linux systems, you can install GPU-enabled packages from conda-forge using the following command:

conda install -c conda-forge implicit implicit-proc=*=gpu

Binary packages are available on conda-forge for Linux, Windows, and OSX. Use the following command to install:

conda install -c conda-forge implicit

To use implicit, ensure you meet the following requirements:

• SciPy: Version 0.16 or later is required.
• GPU Support: Requires RMM to be installed.

The following articles explain the underlying mathematical algorithms used by the library, including Matrix Factorization, GPU acceleration, Approximate Nearest Neighbours, and Distance Metrics:

• Finding Similar Music with Matrix Factorization
• Faster Implicit Matrix Factorization
• Implicit Matrix Factorization on the GPU
• Approximate Nearest Neighbours for Recommender Systems
• Distance Metrics for Fun and Profit

Review these external resources and case studies to see how Implicit is used to build recommendation systems in various contexts:

• H&M Personalized Fashion Recommendations Kaggle Competition
• Yandex Cup 2022: Like Prediction
• Recommending GitHub Repositories with Google BigQuery and the implicit library
• Intro to Implicit Matrix Factorization: Classic ALS with Sketchfab Models
• A Gentle Introduction to Recommender Systems with Implicit Feedback

You can install the implicit package using pip. On x86_64 Linux, Windows, and OSX, this will use prebuilt binary wheels. These wheels include GPU support on Linux for CUDA 13.

pip install implicit

You can install implicit using pip or conda.

• pip: Installs prebuilt binary wheels for x86_64 Linux, Windows, and OSX. On Linux, these wheels include GPU support.
• conda (CPU only): Installs the CPU-only version.
• conda (CPU+GPU): Installs the version with GPU support.

Requirements:

• Python version 3.9 or later.
• SciPy version 0.16 or later.
• For GPU support: NVidia CUDA Toolkit version 13 and rmm-cu13 installed via pip.

You can compare the AlternatingLeastSquares model in implicit against the Spark MLlib implementation using the benchmark_spark.py script. This benchmark compares the Conjugate Gradient solver used in implicit (both CPU and GPU versions) against the Cholesky solver used in Spark.

Prerequisites: To ensure a fair comparison, you must compile Spark with native BLAS support before running the benchmark.

You can use the implicit.datasets.lastfm module to download and load the last.fm 360K dataset into memory. This returns the artist labels, user labels, and a sparse matrix of plays.

Note that artist_user_plays is a scipy sparse matrix where rows are musicians and columns are users. The non-zero entries represent play counts.

from implicit.datasets.lastfm import get_lastfm

artists, users, artist_user_plays = get_lastfm()

To avoid performance degradation caused by nested multithreading, you should disable the internal multithreading of your BLAS library. implicit uses Cython and OpenMP to parallelize training across all available CPU cores.

If your system uses OpenBLAS, set: export OPENBLAS_NUM_THREADS=1

If your system uses Intel MKL, set: export MKL_NUM_THREADS=1